[Impeller] Recycle HostBuffer arena entries only after GPU completion#188965
Merged
auto-submit[bot] merged 7 commits intoJul 10, 2026
Merged
Conversation
github-actions
Bot
added
engine
10 tasks
bdero
marked this pull request as ready for review
July 4, 2026 13:47
![gemini-code-assist[bot]](https://avatars.githubusercontent.com/in/956858?s=60&v=4){kind=small}
Contributor
There was a problem hiding this comment.
Code Review
This pull request introduces a GpuSubmissionTracker to track the completion of submitted command buffers, allowing HostBuffer to safely reuse arena entries without racing active GPU reads. This tracking is integrated across the GLES, Metal, and Vulkan backends, and accompanied by comprehensive unit tests. The review feedback suggests optimizing the GpuSubmissionTracker by replacing std::set with std::vector for tracking pending submissions. Since submission IDs are strictly monotonic, utilizing a sorted vector with std::lower_bound avoids per-frame heap allocations and improves performance.
Contributor
|
The logic here looks good to me, but Jason is probably a better reviewer here. He is OOO until Wednesday - good to wait? |
jason-simmons
previously approved these changes
Jul 8, 2026
| // Sorted, since ids are recorded in increasing order. The pending count | ||
| // tracks GPU queue depth and stays small, so erasure is cheap and steady | ||
| // state performs no heap allocation. | ||
| std::vector<uint64_t> pending_; |
Member
There was a problem hiding this comment.
Add an IPLR_GUARDED_BY(mutex_) annotation here
Member
There was a problem hiding this comment.
This also requires using the impeller::Mutex wrapper class in order to support these annotations on Windows.
Member
Author
jason-simmons
approved these changes
Jul 9, 2026
flutter-dashboard
Bot
removed
the
autosubmit
This was referenced Jul 10, 2026
auto-submit Bot
pushed a commit
to flutter/packages
that referenced
this pull request
Jul 10, 2026
flutter/flutter@dc2a870...f7b66f3 2026-07-10 [email protected] [iOS][test] Fix VSyncClient display link deallocation test on iOS 27 (flutter/flutter#188627) 2026-07-10 [email protected] Fix broken listener code in WindowScope (flutter/flutter#189208) 2026-07-10 [email protected] [web] Preserve text field focus across tab switches (flutter/flutter#188738) 2026-07-10 [email protected] [Impeller] Recycle HostBuffer arena entries only after GPU completion (flutter/flutter#188965) 2026-07-10 [email protected] Assert TextStyle height is not NaN (flutter/flutter#186617) 2026-07-10 [email protected] Roll Fuchsia Test Scripts from dFkTCiDxEtPxYK5Nn... to wLST_A-xfOeGT_5mj... (flutter/flutter#189259) 2026-07-10 [email protected] Roll Dart SDK from a11fb7ed40a5 to 0fc1668c4af4 (5 revisions) (flutter/flutter#189257) 2026-07-10 [email protected] Adds a missing `await` to a `FutureOr` (flutter/flutter#189198) 2026-07-10 [email protected] [web] Repair RenderCanvas CSS size drift (flutter/flutter#188797) 2026-07-10 [email protected] Roll Fuchsia Test Scripts from s5_gZFJ8De9AJalTw... to dFkTCiDxEtPxYK5Nn... (flutter/flutter#189215) 2026-07-10 [email protected] Roll Fuchsia Linux SDK from QcRFUtvCw2EobfJ8s... to czpzDg9ABY2oKLAOY... (flutter/flutter#189222) 2026-07-10 [email protected] Update Depdencies Used By `flutter_engine_group_performance` (flutter/flutter#189229) If this roll has caused a breakage, revert this CL and stop the roller using the controls here: https://autoroll.skia.org/r/flutter-packages Please CC [email protected],[email protected] on the revert to ensure that a human is aware of the problem. To file a bug in Packages: https://github.com/flutter/flutter/issues/new/choose To report a problem with the AutoRoller itself, please file a bug: https://issues.skia.org/issues/new?component=1389291&template=1850622 Documentation for the AutoRoller is here: https://skia.googlesource.com/buildbot/+doc/main/autoroll/README.md
10 tasks
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
3 participants
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
Fixes #184091.
HostBuffer reuses each arena entry kHostBufferArenaSize (4) resets after it was written, assuming no more than 4 frames are ever in flight. Nothing enforces that assumption. On macOS the embedder applies no frame backpressure, so GPU-heavy workloads run the pipeline deeper than the arena and per-frame uniform and vertex data is overwritten while the GPU is still reading it, corrupting rendering.
This makes reuse completion-aware. Each backend context records command buffer submissions and completions in a monotonic watermark, and HostBuffer::Reset retires entries whose GPU work has not yet completed (keeping their buffers alive until it has) and starts them over with fresh allocations. Deep pipelines now cost transient memory instead of correctness, and the memory drains back once the GPU catches up.
Metal tracks completion via command buffer completed handlers and Vulkan via the fence waiter. GLES tracks at reactor consumption granularity, since GL synchronizes buffer reuse implicitly and provides no equivalent completion signal there.
Before fix (repro: #184091 (comment)):
Screen.Recording.2026-07-04.at.2.12.26.AM.mov
After fix:
Screen.Recording.2026-07-04.at.4.38.56.AM.mov
Pre-launch Checklist
///).